Synect
Request access

Everything in Synect, mapped

Strict-factual feature list — drawn directly from FEATURES.md in the source tree. If it's on this page it's shipped; nothing here is on a roadmap.

Chat

A multi-turn chat surface that adapts to whichever model fits the question.

  • Nine providers — OpenAI, Anthropic, Google, Perplexity, Cohere, Groq, ElevenLabs, Runway, HeyGen
  • Auto Mode — two-stage classifier (fast regex rules, then a small-LLM fallback) routes every prompt to the model best-suited for it; explicit override always wins
  • Quality modes — cost_efficient, standard, high_end — switch the whole routing tier at once with org-level enforcement
  • Live web search per provider, with citations to real URLs
  • File attachments (text, PDF, images, audio) with vision-capable models
  • Voice input via Whisper transcription, or full-duplex realtime voice with OpenAI Realtime / Gemini Live
  • Cross-conversation context — pull any previous conversation, label, or whole project into a new prompt as up to 100k chars of grounding
  • Queued messages — line up your next turn while the model is still streaming the current one
  • Regenerate, edit-and-resubmit, branching — every retry creates a versioned branch you can navigate
  • Sharing — public conversation links and embedded share pages
  • Auto-titles, auto-labels, projects, sidebar deep-search across all messages
  • Long-paste detection — pastes >500 chars are auto-converted to attached .md files instead of cluttering the input

Sparks (AI agents)

A Spark is an agent: a system prompt plus a knowledge base plus an optional multi-step pipeline plus opt-in capabilities.

  • System prompt with versioned drafts and live iteration
  • Knowledge base — attach files (PDFs, DOCX, MD, CSV) the Spark always sees
  • Multi-step pipelines — typed steps with routing rules, intermediate state, branching
  • Capabilities — toggle web search, code execution, image generation, deep research per Spark
  • Spark Actions — declarative HTTP / GraphQL actions the Spark can call mid-conversation, with OAuth2 support for Google APIs
  • 80+ Composio integrations — Slack, Notion, Hubspot, Linear, Jira, Salesforce, Stripe, Google Drive/Sheets/Calendar/Docs, Microsoft Teams/Outlook, Zoom, Discord, Asana, Trello, Confluence, Airtable, ClickUp, GitHub, GitLab, Mailchimp, Sendgrid, Twilio, Zoho, Pipedrive, Intercom, and more
  • Custom MCP servers — bring your own Model Context Protocol server (Zapier MCP, Webflow MCP, anything)
  • Marketplace — discover Sparks built by your org or shared publicly
  • End-to-end testing — scripts/test-spark.ts exercises any Spark against Supabase + Composio + the chat route, validates Slides output, runs benchmark comparisons

Canvas

When the model writes a long document, code file, slide deck, or HTML page, Canvas opens alongside the chat — a versioned editor where you can revise the model's draft directly.

  • Streams artifact content live as the model writes
  • Selection-level edits — highlight a paragraph, type an instruction, the model rewrites just that span
  • Full-document rewrites — mode="full" returns a complete updated document
  • Artifact types — markdown, plain text, CSV, HTML, JSON, code (any language with syntax highlighting), PPTX, PDF, DOCX
  • Version history — every Canvas state is snapshotted; navigate back through revisions
  • Round-trip context — when you ask a follow-up, the current Canvas content is injected as context so the model edits in-place rather than starting from scratch
  • Open files from chat — drop a DOCX/CSV/MD into chat and Open in Canvas to start editing with AI
  • DOCX / PDF / PPTX rendering happens server-side (Vercel functions) or in-browser, never via third-party file processors

Media generation

Image, video, music, voice, avatar — all from the same chat input.

  • Image — GPT-Image-1 (OpenAI), Imagen 4 (Google), Flux (Together)
  • Video — Runway Gen-3 / Gen-4 (text-to-video, image-to-video, with extension)
  • Music — ElevenLabs Music (lyrics + composition)
  • Voice — ElevenLabs TTS (natural narration), voice cloning preview
  • Avatar video — HeyGen (talking-head from script + avatar selection)
  • Realtime voice chat — OpenAI Realtime API or Google Gemini Live (full-duplex; same conversation thread)
  • Auto-routing — “generate an image of a hotel lobby” classifies as image_generation and goes straight to GPT-Image-1, no menu hunt

Deep research & web search

When a question needs the web, the right tool fires automatically.

  • Native web search per provider — OpenAI search, Anthropic web search, Google grounding, Perplexity Sonar (built-in)
  • Deep research — Perplexity Sonar Deep Research handles long-running multi-source surveys, streams intermediate findings, harvests citations
  • Page fetch — when you name a specific URL, ScrapingBee retrieves the page HTML for the model to read (gated to OpenAI / Anthropic / Google / Cohere; Perplexity and Groq use their built-in alternatives)
  • SEO data — Sparks with an SEO research step query DataForSEO for keyword volume, competitive analysis, SERP data
  • Citations rendered as superscript [n] links with hover preview and a full reference list at the bottom

Organisations & teams

Multi-user workspace primitives for when you outgrow solo use.

  • Organisations — invite members by email, role-based roles (owner, admin, member)
  • Teams within an org — group permissions, scoped resource access
  • Allowed providers / allowed models — org-level policy that prevents a member from accidentally using an off-policy model
  • Quality-tier policy — restrict members to cost_efficient or standard, reserve high_end for owners
  • Org-level Sparks — Sparks owned by the org, shared automatically with all members
  • Resource sharing — share a project, a label, or a single conversation with another org member
  • Spending controls — per-org token-budget caps with auto-throttle
  • Audit log — pgaudit-backed read/write log of all sensitive table operations

Platform & engineering

What sits underneath the surface — the engineering choices that make Synect dependable.

  • Next.js 16 + React 19 + Turbopack on Vercel; Supabase (Postgres + storage + auth) on AWS
  • RLS-enforced data access — every read and write is scoped to the authed user's tokens, not the service role
  • 14-phase conversation tester — automated coverage of smoke, provider matrix, web search, reasoning, attachments, cross-conversation context, multi-turn dialog, performance/TTFT, media routing, Canvas round-trip, regenerate/branching, Spark chat, governance/error paths, cross-provider web-search XML scrub
  • Server-Timing instrumentation — every chat turn emits per-stage latency breakdown (auth, perms, deps, classifier, sanitize, pre-stream, stream-end)
  • Stream transforms — <think> and <function_calls> XML scrubbers route narrated tool-call leaks to reasoning channels rather than visible text
  • Per-provider tool registration — fetch_page tool stripped for Groq (tool_use_failed retry wrapper) and Perplexity (built-in search active) to prevent narrated XML leaks
  • Edge middleware — 301 redirects from synect.ai/c/* etc. back to app.synect.ai, no broken bookmarks during the domain split

See it in motion

Words on a page only do so much. The fastest way to evaluate Synect is to use it.

Request accessSee pricing